Saurjya Sarkar
Audio ML engineer and researcher in AI & Music
saurjya
London, United Kingdom
Joined October 2025
Network
1.7K connectionsSummary
Academic leader in applied audio source separation: Saurjya's PhD and associated papers focus on time-domain models and permutation-invariant training for monotimbral separation (choirs, chamber ensembles and vocal harmonies). He created/used EnsembleSet (a high-quality synthetic dataset) and demonstrated strong cross-dataset gains via pre-training and fine-tuning, work recognised with the Best Student Paper award at WASPAA 2023. ac+2
Bridges research and industry: He has practical industry experience building audio-quality assessment tools at Qualcomm and holds patents related to voice/immersive audio machine learning. In 2025 he transitioned to an industry audio-ML role at Universal Music Group (Abbey Road), applying research expertise to product-scale music/audio engineering problems. github+2
Publishes open code and datasets: maintains public repositories (e.g., EnsembleSep, EnsembleSet-MIDI) and a personal site with demos and code, supporting reproducible research and enabling others to build on his ensemble / harmony separation work. github+2
Practitioner with music-first perspective: musical background (guitar, music club activity) informs an applied approach to research — designing datasets, augmentations and evaluation metrics that reflect musical structure (e.g., harmonic overlap measures) rather than purely engineering benchmarks. ac+1
Work
Education
Projects
Writing
Time-domain Music Source Separation for Choirs and Ensembles
March 1, 2024PhD thesis presenting time-domain approaches to monotimbral music source separation, introducing EnsembleSet (synthetic dataset), analyses of permutation-invariant training for ensemble/choral separation, and strategies to improve cross-dataset performance.
Leveraging synthetic data for improving chamber ensemble separation
October 1, 2023WASPAA 2023 paper describing a pre-training strategy using synthetic multi-mic EnsembleSet, data augmentation and fine-tuning to improve chamber ensemble separation (Best Student Paper Award at WASPAA 2023).
Vocal Harmony Separation using Time-domain Neural Networks
August 1, 2021Interspeech 2021 paper adapting time-domain encoder-masker-decoder architectures for four-part a cappella vocal separation; includes analysis of harmonic overlap and training strategies for permutation-invariant separation.
Notes
PSPedro Sarmentosaid
SSSaurjya Sarkar
is one of the most impressive people they know working at the intersection of music and technology.Jul 8, 2026